ARTEX and AntiFox: How to identify an AI attack agent in server logs, and what it looks for in configuration files
Open the nginx access log on your server and look for an address sending requests in this pattern: a few requests a second apart, a half-minute pause, then another series adjusted to the responses received. According to Unit 42, this pattern is consistent with an agent waiting for a language model to respond. Reports published in early November covered ARTEX and AntiFox, attack frameworks built around AI agents. If you administer your own Linux server, the useful information in these reports is how the agent behaves, what it learns from, and what it looks for once it gets in. The address lists in the appendices will be out of date within days. The behavior will remain.
Three reports, two frameworks, one week
Censys, November 5. According to Censys ARC, ARTEX is an open-source framework for autonomous penetration testing, powered by a language model. It was published on GitHub in July 2026, is written in Go, and is operated through a web console. A planning agent translates the objective into attack “intents,” while worker agents run tools and return results. The process repeats until the objective is met. The framework works with Anthropic, OpenAI, or any OpenAI-compatible endpoint. According to Censys, CrowdStrike linked ARTEX in October to an attack on financial institutions in South Korea, and Microsoft linked it to Storm-0940's intrusion into ConnectWise customer environments. As of November 4, Censys had counted 1,465 servers with exposed ARTEX consoles. About 650 were in two hosting blocks with identical configurations, leaving about 815 installations that appear to be separate. Censys notes that the console does not need to be internet-accessible, so the total is probably higher. Of these servers, 155 (10.6%) also run other attack tools, and 43 run gateway or reseller software for access to language models. After CrowdStrike's report, the developer announced that the project was moving to closed source. According to Censys, however, archived copies, translations, and versions under other names remain available.
Arctic Wolf, November 10. Arctic Wolf Labs describes AntiFox, a Chinese-language, open-source platform for agent-based penetration testing. According to the report, several recently observed intrusions linked Alibaba Cloud servers hosting AntiFox interfaces to malicious activity against exposed web services and their AWS environments. A web shell was created on a compromised IIS server. It was used to locate and read web.config files and configuration backups, including database connection strings. Configuration files and logs were copied to a web-accessible upload directory. There was also an attempt to connect to an internal production database using credentials extracted from the configuration. The vulnerability that enabled code execution was not identified. In AWS environments, valid identities were used to send hundreds of API calls to various services in about half an hour. Most calls were denied, but some succeeded. In some cases, the activity progressed to the successful retrieval of production credentials from Secrets Manager.
Unit 42, November 11. According to Palo Alto Networks' Unit 42, its incident response team investigated several intrusions in which the attacker's server hosted an ARTEX console or used infrastructure and user agent strings associated with it. As of October, it had counted 609 exposed servers running ARTEX and 1,323 running CyberStrikeAI, a tool in the same family. According to Unit 42, attackers using ARTEX used public password reset pages and exposed APIs to gain cloud access, sometimes in less than a day, without exploiting any software vulnerability. The report also details the behavioral patterns the team identified in ARTEX activity. We will return to those shortly.
TechCrunch's coverage says Censys found more than 800 installations accessible without a password, but Censys's report makes no such claim: 815 is the estimated number of separate installations after excluding the two hosting blocks. TechCrunch also writes that the move to closed source limits future versions to paying customers, while Censys describes an announcement of a move to closed source with no further releases.
The addresses change; the behavior does not
According to Unit 42, ARTEX is continually being forked and modified, and attackers replace their infrastructure quickly. Censys observed that most open directories on ARTEX servers were active for a few hours to a few days: between November 3 and 5, 10 closed and 15 new ones opened. Arctic Wolf warns that public cloud infrastructure can be reassigned to other tenants. An address you copy from an appendix today may belong to a legitimate customer of the same provider next week.
According to Unit 42, the method of operation is more stable than the indicators. Across all its observations, it pointed to an automated planner responding to what it saw. These are the main signs, in our words:
- Think, then act. Bursts of one to 14 requests about a second apart, followed by pauses of 25 to 31 seconds. According to Unit 42, this pattern is consistent with an agent waiting for the model to respond.
- Self-correction. A request that failed with HTTP 422 is corrected and sent again, succeeding moments later.
- Decoy input. Fabricated usernames before real ones, and placeholder values such as
test@example.comordeadbeefdeadbeefdeadbeef12. - Mapping first. Paths from downloaded OpenAPI specifications and source map files are attacked almost immediately.
- Self-checking. A GET request to the same path follows almost every successful write.
- Repeated loops. A request every 48 seconds for three days, and identical attack sequences repeated about 30 seconds apart.
- Machine speed. Moving to the next stage within about a minute of downloading credentials, and checking the validity of more than 200 application credentials (client secrets) in about three minutes.
Unit 42 adds a caveat: any one sign could come from an unusually configured tool or a person testing an API manually. Only the combination points to an autonomous tool. It also notes that people are still involved: in one case, the attacker switched from script-based clients to a desktop browser for the password reset stage and the first token request.
Administrators who run their own servers have an advantage: all these signs are already in their logs. You do not need external intelligence to measure the intervals between requests from the same address.
The agent learns from your responses
According to Unit 42, these tools use the target's responses as context for the next step. That makes them effective, but it is also a weakness. Unit 42 assesses that defenders may be able to exploit this: when the target returns identical, generic error codes or quietly slows connections, the agent has no clues to help it adjust its next step. It wastes context window space and compute budget on a dead end.
On your server, this means checking which responses give the agent information: a message that distinguishes “user does not exist” from “incorrect password”; an input validation error that specifies exactly which field is missing and the required format; an endpoint that checks whether an account exists without requiring authentication; an openapi.json file or source map left in the web root from development. Each supplies information the agent uses to build its next request. This is exactly what the “self-correction” pattern described by Unit 42 looks like: a 422 error, a correction, then success.
You do not need to hide information from legitimate users everywhere. Focus on public paths that do not require authentication: registration, password resets, verification codes, and externally accessible APIs.
From foothold to credentials: minutes, not hours
The first steps on the compromised server were to establish the context and permissions. Next came a search for configuration files, reading connection strings and application settings, and attempts to use the credentials found. According to Arctic Wolf, if a compromised application stored AWS credentials, SDK settings, or references to Secrets Manager, malicious AWS activity may begin shortly after the configuration is discovered. Unit 42 describes a move to the next stage within about a minute of downloading credentials.
Unit 42's conclusion is clear: autonomous tools perform checks faster than people can review alerts. Defenses therefore need to deny access by default and keep the credentials these tools seek out of reach. If a file containing credentials is accessible to the web server process, it is also accessible to anyone who gains the ability to execute code through that process.
Arctic Wolf makes another point relevant to reading logs: do not assume a sequence of actions failed just because most requests were denied. If your log contains many 403 and 404 responses for the same address, with a few 200 responses among them, examine the 200 responses in particular.
Your server can also be on the other side
Censys emphasizes that not every ARTEX installation is malicious, and some exposed servers are probably operated by authorized penetration testers. Unit 42 even recommends that defenders run the same open-source frameworks in their own environments to find exposures before external agents do. If you do this, the reports show what can go wrong. The ARTEX console does not need to be internet-accessible, yet Censys found 1,465 exposed consoles. Its default port is 8787, which Arctic Wolf says is also the common port for AntiFox. ARTEX operators themselves left directories open. According to Censys, one exposed an entire home directory, including .ssh and .bash_history. Censys adds that temporary Python SimpleHTTP servers on high ports often appear before the console does. TechCrunch notes that an exposed console could allow outsiders to run the tool against other targets.
The same reasoning applies to model access gateways. According to Censys, 43 ARTEX servers ran gateway or reseller software for access to language models, mainly CLIProxyAPI and New API. It recommends investigating unfamiliar gateways and resellers that appear in outbound traffic logs. If you run such a gateway, make sure it is not open to everyone and that you know where it connects.
Identify an agent by its timing
- Intervals by address. The following script (gawk) reads an nginx access log in combined format. For each address, it counts pauses of 25 to 31 seconds between requests and requests arriving no more than two seconds apart. An address with high counts in both categories warrants investigation, not automatic blocking:
sudo gawk '{
split($4, t, /[\[\/:]/)
m = (index("JanFebMarAprMayJunJulAugSepOctNovDec", t[3]) + 2) / 3
ts = mktime(t[4] " " m " " t[2] " " t[5] " " t[6] " " t[7])
ip = $1
if (ip in last) { g = ts - last[ip]; if (g >= 25 && g <= 31) p[ip]++; else if (g <= 2) b[ip]++ }
last[ip] = ts
} END { for (ip in p) if (p[ip] >= 5 && b[ip] >= 5) print p[ip], b[ip], ip }' /var/log/nginx/access.log | sort -rn | head -20
- Decoy values.
sudo grep -E 'deadbeefdeadbeef|test(@|%40)example' /var/log/nginx/access.log | awk '{print $1}' | sort | uniq -c | sort -rn. The log contains only the query string, not POST request bodies, so this is a partial check. - Fabricated usernames in SSH.
sudo grep -h 'Invalid user' /var/log/auth.log | awk '{print $(NF-2)}' | sort | uniq -c | sort -rn | head. An address that tries dozens of nonexistent usernames before a real one fits Unit 42's description. - 422 followed by success.
sudo awk '$9 == 422 {print $1, $7}' /var/log/nginx/access.log | sort | uniq -c | sort -rn | head. Then check whether the same address received a 2xx response on the same path seconds later. - Successes among denials. For a suspicious address, filter for successful responses only:
sudo awk -v ip=203.0.113.7 '$1 == ip && $9 ~ /^2/' /var/log/nginx/access.log.
Give it nothing to learn from
- Consistent responses in public flows. For login, registration, and password resets, return the same response whether the account exists or not. Remove endpoints that check whether an account exists without requiring authentication, as Unit 42 recommends.
- Quiet delays. In nginx,
limit_reqwithoutnodelaydelays requests within the burst allowance instead of rejecting them. The client receives a slow response rather than a detailed error message. Beyond the burst allowance, nginx always returns the same error code. The paths here are examples only:
# Inside the http block
limit_req_zone $binary_remote_addr zone=authflows:10m rate=10r/m;
# Inside the server block
location ~ ^/(login|register|password-reset) {
limit_req zone=authflows burst=5;
# ... Existing settings for these paths
}
- Require authentication on every endpoint that performs writes, including development and staging interfaces, as Unit 42 recommends.
- Remove mapping files. Remove OpenAPI specifications, source maps, and configuration files containing tokens or client IDs from externally accessible applications.
Keep credentials out of reach
- Look for files that should not be in the web root:
sudo find /var/www -type f \( -name '.env*' -o -name '*.bak' -o -name '*.old' -o -name '*.sql' -o -name '*.map' -o -iname '*openapi*' -o -iname '*swagger*' \) 2>/dev/null
- Search for credential strings:
sudo grep -rIlE --exclude-dir=node_modules 'aws_secret_access_key|AKIA[0-9A-Z]{16}|DB_PASSWORD|DATABASE_URL' /var/www 2>/dev/null. - Move configuration files outside the web root and restrict access to the application's user account. To find files readable by any user:
sudo find /var/www -type f -perm -o=r \( -name '.env*' -o -name '*config*' \). - Check upload directories. Arctic Wolf observed configuration files copied to an upload directory with a
.txtextension. Check for new files in these directories:sudo find /var/www -path '*upload*' -type f \( -name '*.php*' -o -name '*.phtml' -o -name '*.txt' \) -mtime -14. - On a cloud server, replace long-lived access keys with role-based permissions, as Unit 42 recommends. Following Arctic Wolf's guidance, enforce IMDSv2 with a restrictive hop limit and give application identities only the permissions they need.
- Usage alerts. Unit 42 recommends looking for the first use of a cloud access key from a new IP address. Arctic Wolf recommends alerting on a successful GetSecretValue call from new infrastructure, especially after a series of denials.
- If anything was exposed, rotate database passwords, cloud access keys, API tokens, and encryption keys. Arctic Wolf recommends first revoking active sessions and keys and reviewing each identity's activity, and only then issuing replacements.
Make sure your server is not the console
- Console ports:
sudo ss -ltnp | grep -E ':(8787|8790) '. - Forgotten file servers:
ps -eo pid,user,args | grep -E '[h]ttp\.server|[S]impleHTTP'. - If you run a penetration testing framework or a model gateway, configure it to listen on 127.0.0.1 and connect through an SSH tunnel. Check where the gateway sends its requests.
- Do not expose a home directory through a web server, or any directory containing
.sshor shell history, even for a few minutes.
How nuDefend helps
What the server exposes, including its AI tools. nudefend expose shows every listening service on the server, including services in containers, and its access status: local only, private network, blocked by a local rule, externally accessible, or published from a container. An externally accessible service that is unexpected on its listening port is flagged for review. When a previously internal service becomes externally accessible, nuDefend alerts you. nudefend workloads shows the AI tools running on the server, on the host or in a container: model servers, chat interfaces, vector databases, automation tools, agent platforms, and MCP servers. Each entry includes the information used to identify it. You can close a port that should not be open from the dashboard or with nudefend block add, after a preview showing exactly what will be blocked.
Where AI tools connect. nuDefend records the destinations each AI tool on the server connects to, how much traffic was sent to each destination, and when the destination was first detected. When a destination is identified, its name and type of activity are shown, such as a model service through which requests and responses pass. An unidentified destination is marked as such, making it stand out in the list. On the Outbound screen, you approve the destinations the tool actually needs. Once the restriction is enabled, the tool can connect only to those destinations. Any attempt to connect elsewhere is blocked before traffic leaves the server, and a restricted tool can resolve only addresses you have approved. The rest of the server continues to operate normally.
Password guessing and login page scanning. nuDefend reads login logs for SSH, mail, FTP, and the login pages of WordPress, phpMyAdmin, Joomla, and Drupal. It blocks an address at the firewall if it exceeds the threshold within a 10-minute window. Three requests for a login page that does not exist on the server count as a scan, and the address is blocked.
Threat blocking in both directions. A cryptographically signed threat list blocks traffic to and from known malicious addresses at the kernel firewall, including active scanners, sources of password-guessing attempts, and botnet command-and-control servers. Unlike a list copied from a report appendix, it updates automatically every 30 minutes. At the DNS layer, dangerous domains are blocked system-wide, so no application on the server can resolve or reach them. If anything on the server tries to contact a malicious address, the blocked outbound traffic card on the dashboard turns red as an early warning. All reports remain on the server.
More details are available on the exposure page, the outbound traffic page, the access blocking page, and in the documentation.
The bottom line
The appendices of these three reports will be out of date by the end of the month. The agent will not change at the same pace: it will send a series of requests, wait for the model, correct itself based on the response, and look for a configuration file containing credentials as soon as it gets in. When you administer your own server, its protection is in your hands: measure request timing in the logs, return responses on public paths that give the agent no useful information, keep credentials out of the web server's reach, and check that your server is not itself running an attack console or model gateway open to everyone.
nuDefend shows every listening service on the server and who can reach it, including AI tools and services in containers, and alerts you when a previously internal service becomes externally accessible.